I believe it's changed alot since then? I was abit cynical of it from that initial experience but recent usage of it made it seem reliable and scalable.
Until storage is solved (and it might not ever be), then nuclear is excellent for base load. Germany lacks coast line for wind and is in northern europe. Nuclear is ideal for them.
Caching in this scenario isn't something I'd lean on unless you can invalidate or repopulate easily. I've used etcd in a feature flag scenario due to the speed of it's replication and it's ability to queried frequently without the need for caching.
I worked with a mesos config app, backed by git. Watching the whole go down when github died resolved none of us to ever trust it again for that sort of role.
It isn't a permanent solution, it's purely designed to drop weight. You have to do the mental and physical work to be able to maintain which is a different mindset.
It's not about everything fits, it's just much easier for orgs to have a consistent environment for developers to use. Google does this to great success, getting out of the box logging, metrics, a universal way to define a system, access secrets. A whole class of decisions stop being made all the time. Anyone in the org can look in a repo and understand what and where everything is deployed. It's simple stuff like this that make it compelling.
I lol'd when I got to 'Why would you ever want to allocate more resources to one particular part? It’s not like the other parts will eat up the extra resources. If your system needs more RAM, it needs more RAM. Why would you care about which part needs more RAM?' and stopped reading.